The differences between MCP, Function Calling, A2A, and Skills
"What's the difference between MCP and Function Calling" and "What's the difference between MCP and A2A" are high-volume search queries. After finishing the previous sections, you've actually already used all of them. This section puts them together in one diagram.
Mapping them with myagent
| Concept | Between whom and whom | Where it sits in myagent |
|---|---|---|
| Function Calling | model ↔ your program | tools=... parameters, and the tool_calls coming back from the model |
| MCP | your program (host) ↔ external tool server | --mcp connecting to notes_server.py |
| Agent Skills | instructions for how to do things, written for the model | skills/weekly-report/SKILL.md |
| A2A | one Agent ↔ another Agent | not used in myagent |
Function Calling and MCP: two layers, not an either-or choice
In our code, a single MCP tool call goes through two layers:
- 1?
- 2?
- 3?
- 4?
- 5?
Click them one by one, in order.
So:
- Function Calling answers "how does the model express 'I want to call a certain function'". Each model provider's syntax differs slightly (for example, OpenAI's newer Responses API and the Chat Completions format are different).
- MCP answers "where are tools written, and how are they discovered and invoked". Write a tool once, and any MCP-supporting application can use it, regardless of which model is behind it.
MCP and A2A: vertical and horizontal
The A2A (Agent2Agent) protocol's official comparison puts it clearly:
- MCP is vertical: connecting a single Agent to more tools and data, making it deeper.
- A2A is horizontal: letting different Agents collaborate with each other—they may come from different companies, use different frameworks, and run on different servers.
A few key points about A2A: it's based on JSON-RPC 2.0 over HTTP(S); each Agent uses an "Agent Card" to describe its capabilities for others to discover; it supports synchronous, streaming, and asynchronous push; the other Agent's internal state, memory, and tools are not exposed—you treat it as a colleague who can take on tasks, not as a tool. A2A was contributed by Google, the v1.0 specification was released in March 2026 (month 3), and in August (month 8) it joined the AAIF under the Linux Foundation, becoming a sibling project to MCP. The Python SDK is pip install a2a-sdk.
The two are complementary: each Agent uses MCP to connect its own tools, while Agents collaborate with each other via A2A.
Skills: the "how-to" written for the model
Skills are not a communication protocol; they're a file format: a folder with a SKILL.md that tells the model how to handle a certain class of tasks. They don't conflict with MCP either: MCP provides capabilities, and a Skill explains how to use those capabilities to get something done. (AAIF and Microsoft materials also describe an extension for distributing skills via MCP, which is a fairly new approach.)
One diagram to remember it
Look it upI want to...
Click one to see.
Quiz time“Will MCP replace Function Calling?”
Check yourself
- I can state which two parties each of the four concepts sits between
- I can explain why MCP and Function Calling are two layers rather than an either-or choice
- I can distinguish MCP and A2A using "vertical vs. horizontal"
Next chapter: how to choose a framework.